A clarification of what userjourneys is building. Written for those who might want to join us.
text 40 bits/s
....................................................
: :
: screen v
+--------+ 1.6 Gbit/s +-------------+ +-------------+
| AI |==============>| OUTER BRAIN |------->| INNER BRAIN |
| | | parallel | | serial |
+--------+ +-------------+ +-------------+
^ |
+------------------- 10 bits/s <-------------------+
Typing at 120 words a minute produces about ten bits a second, as English carries ~ one bit per character. In fact, in any single domain, where peak human cognitive performance has been measured, it always lands at ten bits per second. A StarCraft professional at a thousand actions a minute lands at the same number, because most of those actions are redundant. A blindfolded speedcuber memorizing a cube manages 12 bits a second. A world champion memorizing binary digits manages five.
Meanwhile the senses take in about a billion bits per second. The cones in one eye alone deliver 1.6 gigabits. The gap between what comes in and what we do with it is a factor of a hundred million, and nobody really knows why.
It has been proposed that the brain runs in two modes. An outer brain that talks to the world: millions of sensors and muscles, massively parallel, extremely fast. And an inner brain that decides: a reduced stream of a few bits, strictly serial, one thought at a time. Everything you would call high-level cognition happens in the inner brain, at ten bits per second, and there is no known way to speed it up.
If the inner brain is fixed at ten bits a second, then making someone more capable becomes a game of spending those ten bits as efficiently as possible. A person should use them on the decisions that only they can make and that matter most, everything else should be left to AI.
For the AI to take work off the inner brain, it needs a model of the person and of what they want. It can build that model from a few sources.
Once the AI has modelled intent as well as the available information allows, it has to act. It has to make decisions under uncertainty. And often it will hit a point where it should not guess. It has to present options, or ask a clarifying question, and the person chooses what happens next. That choice is where the ten bits get spent. So how do you ask as efficiently as possible?
There are only a few ways to get options in front of a person.
Write them straight into the brain. This seems to be the strongest possible answer. Cochlear implants have written into the auditory nerve for forty years, and Neuralink is starting trials of an implant that stimulates the visual cortex. Both restore a broken sense at very (very) low resolution. Putting signal directly into the brain is one of the most outlandish ideas in sci-fi and happens deep into the 21st century on most people's timelines.
Go through the senses. This is the channel we have to build on for the coming decades, and we have three modalities to choose between: text, audio and visuals.
| Channel | Intake |
|---|---|
| Reading text | 28 - 45 bits per second |
| Listening to speech | 13 - 39 bits per second |
| Looking at an object | 30 - 50 bits per second |
A word and a drawn object arrive at about the same rate, 30 to 50 bits per second. If "a picture is worth a thousand words" meant that pictures carry more bits into your head, it would be false.
But using pictures to communicate bits of information to a brain comes with a clever trick we can use. The visual cortex takes in the entire visual scene in front of you at once and computes, in parallel, where the thing you care about is. This massive parallel computation is extraordinarily cheap, as it never touches the inner brain. Each glance is a query to that parallel engine that decides where you should look next.
Text cannot be searched using this massively parallelized engine. It is not possible to know that a sentence is irrelevant until you have read it. So the cost of a paragraph is linear in its length, and the cost of a well-built screen is logarithmic in the number of options: one glance per level of hierarchy.
What this means is that we can ask a person every question a text could ask, all at once, and let their visual system find the one that matters. The inner brain then spends its ten bits on that one. Prose that takes minutes to read can be answered in seconds as a picture.
You are booking a flight and you need a seat. Here is the availability of seats on the flight as text:
Window seats are open in rows 3, 5 and 6, but each of them has a stranger in the middle seat. Row 9 has an empty set of three in the middle. On the left side near the toilet, in row 12, there is an aisle seat with nobody in the middle. That one is $50, the others are $500 and $20. There is also...
Here is a visual representation of the same situation:
The text takes minutes to read and leads to a high likelihood of mistakes, as you have to hold a lot of state like "row 9, middle, $20" in your head while you keep reading. Meanwhile, the visual question takes seconds to answer. Both versions of the question ask for the same eight bits from the user that define which seat they want.
Every AI product today clarifies intent through the narrow channel, text and reading. Every clarification goes through the serial pipe of the inner brain at reading speed, and the widest channel into the brain is used to show a chat transcript.
Today AI clarifies intent through text, at a cost that grows linearly with the number of options. A visual interface makes that cost logarithmic. Taking clarification from O(n) to O(log n) is the bottleneck we are working on.
The visual interfaces do not have to look the same for everyone. They have to be consistent for you. Every interaction with the digital world should be rendered into your own visual language, and we are training the model that does that.
Sources